CN112513891A

CN112513891A - Information processing apparatus, information processing method, and program

Info

Publication number: CN112513891A
Application number: CN201980049686.4A
Authority: CN
Inventors: 藤村亮太; 中田洋平; 筑泽宗太郎
Original assignee: Panasonic Intellectual Property Corp of America
Current assignee: Panasonic Intellectual Property Corp of America
Priority date: 2019-01-02
Filing date: 2019-12-10
Publication date: 2021-03-16
Also published as: EP3907670A4; JP7483631B2; EP3907670A1; WO2020141587A1; US20210166139A1; JPWO2020141587A1

Abstract

An information processing device (100a) is an information processing device provided with a processor that acquires a 1 st classification threshold value for classifying data into at least 1 of a plurality of classes; outputting a classification result that classifies the data into at least 1 class of the plurality of classes based on an output of the trained classification model and a 1 st classification threshold; the 1 st classification threshold is obtained by a 2 nd transform of the 2 nd classification threshold, the 2 nd transform being an inverse transform to the 1 st transform; the 1 st transformation is a transformation of classification probability values from the output of the trained classification model to a plurality of individual classes constituting a plurality of classes; the 2 nd classification threshold is set based on classification probability values of the plurality of monomer classes.

Description

Information processing apparatus, information processing method, and program

Technical Field

The invention relates to an information processing apparatus, an information processing method and a program.

Background

In the field of image recognition and the like, a multi-layer Neural Network (DNN) is used as a recognition model (hereinafter, also referred to as a classification model) for recognizing an object in an image. The DNN takes an image as an input, for example, and outputs a probability value (also referred to as a likelihood for an object class) of classification of a class of an object included in the image. At this time, a Softmax function is used in the output layer of DNN (for example, see patent document 1).

Documents of the prior art

Patent document

Patent document 1: international publication No. 2017/149722

Disclosure of Invention

Problems to be solved by the invention

However, since the Softmax function includes an exponential operation, there is a possibility that the computational resources are urgent when the classification model is installed in an embedded device in which computational resources are limited.

Therefore, the present invention provides an information processing device, an information processing method, and a program that can reduce the amount of computation for classifying categories of objects.

Means for solving the problems

In order to solve the above problem, an information processing apparatus according to an aspect of the present invention is an information processing apparatus including a processor that acquires a 1 st classification threshold value for classifying data into at least 1 of a plurality of classes; outputting a classification result of classifying the data into at least 1 of the plurality of classes based on an output of a trained classification model and the 1 st classification threshold; the 1 st classification threshold is obtained by a 2 nd transform of a 2 nd classification threshold, the 2 nd transform being an inverse transform to the 1 st transform; the 1 st transformation is a transformation of classification probability values from an output of the trained classification model to a plurality of individual classes constituting the plurality of classes; the 2 nd classification threshold is set based on the classification probability values of the plurality of cell types.

In addition, an information processing method according to an aspect of the present invention is an information processing method executed by a computer, and performs 1 st conversion of classification probability values from an output of a trained classification model to a plurality of cell classes; setting a 2 nd classification threshold value based on the classification probability values of the plurality of monomer classes; performing a 2 nd transform, said 2 nd transform being a transform from the above 2 nd classification threshold to a 1 st classification threshold used to classify the data into at least 1 of the plurality of classes, being an inverse transform to the 1 st transform; and outputting the 1 st classification threshold value.

Further, a program according to an aspect of the present invention causes a computer to execute an information processing method including: obtaining a 1 st classification threshold for classifying the data into at least 1 of a plurality of classes; outputting a classification result of classifying the data into at least 1 of the plurality of classes based on an output of a trained classification model and the 1 st classification threshold; the 1 st classification threshold is obtained by a 2 nd transform of a 2 nd classification threshold, the 2 nd transform being an inverse transform to the 1 st transform; the 1 st transformation is a transformation of classification probability values from an output of the trained classification model to a plurality of individual classes constituting the plurality of classes; the 2 nd classification threshold is set based on the classification probability values of the plurality of cell types. Alternatively, an aspect of the present invention can be realized as a computer-readable non-transitory recording medium storing the program.

Effects of the invention

According to the present invention, the amount of computation for classifying the types of objects can be reduced.

Drawings

Fig. 1 is a block diagram showing an example of the configuration of an information processing system according to embodiment 1.

Fig. 2 is a flowchart showing an example of the operation of the threshold value calculation device according to embodiment 1.

Fig. 3 is a flowchart showing an example of the operation of the information processing apparatus according to embodiment 1.

Fig. 4 is a block diagram showing an example of the configuration of an information processing apparatus according to a modification of embodiment 1.

Fig. 5 is a flowchart showing an example of the operation of the threshold value calculation unit according to the modification of embodiment 1.

Fig. 6 is a flowchart showing an example of the operation of the information processing unit according to the modification of embodiment 1.

Fig. 7 is a block diagram showing an example of the configuration of the information processing system according to embodiment 2.

Fig. 8 is a flowchart showing an example of the operation of the threshold value calculation device according to embodiment 2.

Fig. 9 is a block diagram showing an example of the configuration of an information processing apparatus according to a modification of embodiment 2.

Fig. 10 is a flowchart showing an example of the operation of the threshold value calculation unit according to the modification of embodiment 2.

Detailed Description

(reaching the knowledge of the present invention)

In the past, in a multilayer Neural Network (DNN) which is mounted in a computing device with limited computing resources such as an embedded device, the number of units of a hidden layer cannot be increased, and thus, the problem of the degradation of the pattern recognition performance is caused. To solve this problem, the conventional technique described in patent document 1 determines whether or not scalar quantization (scalar quantization) is performed for each DNN layer, and multiplies a vector subjected to scalar quantization by a weight vector in a layer next to the layer subjected to scalar quantization. Thus, the number of units of the hidden layer can be increased because the amount of calculation can be reduced as compared with the case of multiplying a vector that is not scalar-quantized by a weight vector. However, since the likelihood vector is calculated using the output vector that is scalar-quantized, the value becomes coarse compared to the case where the likelihood vector is calculated using the output vector that is not scalar-quantized. Therefore, the recognition accuracy may be degraded.

In the conventional technique described in patent document 1, when scalar quantization is not performed on the first 1 layers of the output layer of DNN, a Softmax function is applied to the output layer to calculate likelihood vectors (hereinafter referred to as classification probability values) for a plurality of classes. Since the Softmax function includes the operation of the exponential function, the amount of operation of the exponential function becomes a problem when the Softmax function is installed in the embedded system. Furthermore, the Softmax function adjusts the input so that the sum becomes 1, and therefore, the original value is not restored even by the inverse transform. In other words, since the Softmax function performs an irreversible operation on the input, it is impossible to obtain a value that is not normalized by inversely transforming the output value obtained by inputting the input to the Softmax function. Therefore, in the conventional technique described in patent literature 1, for example, in order to recognize an object in an input image, it is necessary to calculate classification probability values for each of a plurality of categories of the object in the input image. In the process of calculating the classification probability value of the object in the input image, the calculation amount of the calculation device with limited calculation resources such as the embedded device is increased, and therefore, the DNN recognition accuracy mounted on the calculation device may be lowered. Therefore, it is difficult to say that the conventional technique described in patent document 1 can reduce the amount of computation for classifying the types of objects.

The present inventors have made extensive studies in view of the above-described problems, and as a result, have found that, in the process of determining the threshold values for each of the plurality of types, by performing reversible transformation on the output layer of the DNN, the calculated threshold values can be inversely transformed, and threshold values that are not normalized can be obtained. Thus, for example, a threshold value that is not standardized can be used in the classification processing of the object in the input image, and therefore an information processing device that can reduce the amount of computation for classifying the object is conceived.

An outline of an embodiment of the present invention is as follows.

An information processing apparatus according to an aspect of the present invention is an information processing apparatus including a processor that acquires a 1 st classification threshold value for classifying data into at least 1 category of a plurality of categories; outputting a classification result of classifying the data into at least 1 of the plurality of classes based on an output of a trained classification model and the 1 st classification threshold; the 1 st classification threshold is obtained by a 2 nd transform of a 2 nd classification threshold, the 2 nd transform being an inverse transform to the 1 st transform; the 1 st transformation is a transformation of classification probability values from an output of the trained classification model to a plurality of individual classes constituting the plurality of classes; the 2 nd classification threshold is set based on the classification probability values of the plurality of cell types.

According to the above configuration, in the 1 st transformation for transforming the output of the classification model into the classification probability values of the plurality of cell classes, a function of the reversible transformation is used. Therefore, if the 1 st classification threshold obtained by performing the 2 nd transform, which is an inverse transform to the 1 st transform, on the 2 nd classification threshold is used in the process of classifying the classes of the object, it is no longer necessary to transform, for example, the output of the classification model with the image as an input into the classification probability values of the plurality of individual classes. Therefore, according to the information processing apparatus pertaining to one aspect of the present invention, the amount of computation for classifying the types of objects can be reduced.

Specifically, in the information processing apparatus according to an aspect of the present invention, the output of the trained classification model may be a plurality of scalars corresponding to the plurality of classes.

According to the above method, in the 1 st transformation from the output of the classification model to the classification probability values of the plurality of cell classes, a function of the reversible transformation is used. Therefore, by performing the 2 nd transform, which is an inverse transform to the 1 st transform, on the 2 nd classification threshold, the 1 st classification threshold, which is a threshold not normalized, is obtained. For example, if the 1 st classification threshold is used in the process of classifying the class of the object, it is no longer necessary to transform the output of the classification model into classification probability values for a plurality of individual classes. Therefore, according to the information processing method pertaining to one aspect of the present invention, since the 1 st classification threshold that is a threshold that is not normalized can be obtained, the amount of computation for classifying the class of the object can be reduced.

For example, in the information processing method according to one aspect of the present invention, the 1 st transformation may be an operation based on a probability function that can be inversely transformed; the 2 nd transformation is an operation based on an inverse function of the probability function.

By this means, by inversely transforming the classification probability values of the plurality of individual classes, it is possible to derive the values before transformation, i.e., the output of the trained classification model.

For example, in the information processing method according to one aspect of the present invention, the 1 st transformation may be a transformation using a database corresponding to an operation based on a probability function that can be inversely transformed; the 2 nd transformation is a transformation using a database corresponding to an operation based on an inverse function of the probability function.

This can further reduce the amount of calculation compared to the function operation.

For example, in the information processing method according to an aspect of the present invention, a data set may be acquired; inputting the data set to the trained classification model, and obtaining classification probability values of the plurality of cell types for each data included in the data set; the 2 nd classification threshold is determined based on the classification result using the 2 nd classification threshold for each of the acquired classification probability values of the plurality of cell classes.

Thus, the 2 nd classification threshold is determined by referring to the correct solution data included in the evaluation data set, changing the value of the 2 nd classification threshold, which is the threshold of the classification probability values for the plurality of individual categories, and selecting the 2 nd classification threshold that satisfies the target accuracy. Therefore, according to the information processing method pertaining to one aspect of the present invention, it is possible to determine the threshold value at which a desired classification accuracy can be obtained.

Further, a program according to an aspect of the present invention causes a computer to execute an information processing method including: obtaining a 1 st classification threshold for classifying the data into at least 1 of a plurality of classes; outputting a classification result of classifying the data into at least 1 of the plurality of classes based on an output of a trained classification model and the 1 st classification threshold; the 1 st classification threshold is obtained by a 2 nd transform of a 2 nd classification threshold, the 2 nd transform being an inverse transform to the 1 st transform; the 1 st transformation is a transformation of classification probability values from an output of the trained classification model to a plurality of individual classes constituting the plurality of classes; the 2 nd classification threshold is set based on the classification probability values of the plurality of cell types.

According to the above-described procedure, in the 1 st transformation from the output of the classification model to the classification probability values of the plurality of cell classes, a function of the reversible transformation is used. Therefore, if the 1 st classification threshold obtained by performing the 2 nd transform, which is an inverse transform to the 1 st transform, on the 2 nd classification threshold is used in the process of classifying the classes of the object, it is no longer necessary to transform, for example, the output of the classification model with the image as an input into the classification probability values of the plurality of individual classes. Therefore, according to the program pertaining to one aspect of the present invention, the amount of computation for classifying the types of objects can be reduced.

Hereinafter, the embodiments will be specifically described with reference to the drawings.

The embodiments described below are all illustrative or specific examples. Therefore, the numerical values, shapes, components, arrangement positions and connection forms of the components, steps, order of the steps, and the like shown in the following embodiments are examples, and do not limit the present invention. Further, among the components of the following embodiments, components that are not recited in the independent claims are described as arbitrary components.

The drawings are schematic and not necessarily strictly illustrated. Therefore, for example, the scales and the like do not always match in each drawing. In the drawings, substantially the same components are denoted by the same reference numerals, and redundant description is omitted or simplified.

In the present specification, terms and numerical ranges indicating relationships between elements such as horizontal and vertical are not words of strict meanings, but words including substantially equivalent ranges, for example, differences of about several percent.

(embodiment mode 1)

[ overview of information processing System ]

First, an outline of an information processing system including an information processing device according to embodiment 1 will be described with reference to the drawings. Fig. 1 is a block diagram showing an example of the configuration of an information processing system 300a according to embodiment 1.

The information processing system 300a is a system that classifies data acquired by a sensor into at least 1 of a plurality of classes and outputs the classification result. The information processing system 300a includes: a threshold calculation means 200a that calculates a 1 st classification threshold used to classify the data into at least 1 of the plurality of classes; and an information processing device 100a that outputs a classification result of classifying the data into at least 1 of the plurality of classes based on the output of the trained classification model and the 1 st classification threshold.

The sensor is, for example, a sound sensor such as a microphone, an image sensor, a distance sensor, a gyro sensor, a pressure sensor, or the like. Data acquired using a plurality of sensors may also be acquired using a three-dimensional reconstruction technique such as SfM (Motion from Motion). The data acquired by the sensor is, for example, sound, image, moving image, three-Dimensional point group (three-Dimensional point group) data, or vector data.

In the information processing system 300a, a plurality of categories into which data is classified may be set according to the type of data, the use of data, and the like. For example, when the data is voice, the type of voice of a specific person, a specific mechanical operation sound, a specific animal sound, or the like may be set. In the case where the data is an image, for example, in a monitoring camera system or the like, the type of a specific person or the like may be set, and in an in-vehicle camera system or the like, the type of a pedestrian, a car, a motorcycle, a bicycle, a background or the like may be set. When the data is three-dimensional point group data, the type of the structure or the topography may be set according to the three-dimensional shape of the structure or the topography, the unevenness, the crack, the specific structure, or the like, for example. In the case where the data is vector data, for example, the types of motion vectors and the like of a plurality of portions of a structure such as a bridge or a sound insulating wall may be set.

The following describes each configuration of the information processing system 300 a.

[ threshold value calculating means ]

The threshold calculation means 200a is means for calculating the 1 st classification threshold used to classify data into at least 1 of a plurality of classes.

As shown in fig. 1, the threshold calculation device 200a includes a storage unit 201, a 1 st calculation unit 202, a classification probability calculation unit 203, a classification threshold determination unit 204, a threshold conversion unit 205, and a 1 st output unit 206.

The storage unit 201 stores the evaluation data set of the 2 nd classification threshold. The evaluation data set includes a set of input data input to the 1 st arithmetic unit 202 and correct solution data corresponding to the input data. The positive solution data is a classification probability value for each of a plurality of classes of the input data. Hereinafter, the classification probability values for each of the plurality of classes will also be referred to as classification probability values for the plurality of individual classes. The monomer type means each of a plurality of types. The classification probability values of the plurality of cell categories are normalized probability values obtained by subjecting the output of the 1 st arithmetic unit 202 to the 1 st conversion in the classification probability arithmetic unit 203. The 2 nd classification threshold is set based on classification probability values of the plurality of monomer classes.

The 1 st arithmetic unit 202 is a feature extractor for extracting a feature of data, and is, for example, a machine learning model. For example, the 1 st arithmetic unit 202 is a trained classification model. The classification model is a multi-layer neural network (DNN). The 1 st arithmetic unit 202 acquires an evaluation dataset. For example, the 1 st arithmetic unit 202 reads out the evaluation data set from the storage unit 201. The input data of the evaluation data set is input to the 1 st arithmetic unit 202. The 1 st arithmetic unit 202 outputs a plurality of scalars corresponding to a plurality of categories of the input data. Each scalar quantity is a feature quantity corresponding to each category of the input data. The output of the 1 st arithmetic unit 202 is a value that is not normalized.

The 1 st arithmetic unit 202 is not limited to DNN. For example, the 1 st arithmetic unit 202 may be a feature extractor other than the DNN that uses a method such as an edge extraction method, a principal component analysis method, a block matching method, or a sampling moll method.

The 1 st arithmetic unit 202 may acquire the evaluation dataset from another device via communication. For example, the evaluation dataset may be acquired from a server or a storage device via the internet.

The classification probability calculation unit 203 performs 1 st conversion, which is conversion of the output from the 1 st calculation unit 202 into classification probability values of a plurality of cell categories. More specifically, the classification probability calculation unit 203 calculates classification probability values of a plurality of cell classes from the output of the 1 st calculation unit 202 using a reversible transformation in the 1 st transformation. For example, the classification probability calculation unit 203 normalizes a plurality of scalars (feature quantities) of the input data corresponding to a plurality of classes using a probability function of reversible transformation, thereby deriving classification probability values of the input data corresponding to the plurality of classes. By normalizing by the reversible transformation in this way, it is possible to determine an appropriate threshold value (the 2 nd classification threshold value described later) based on the evaluation data set, and then, inversely transform the appropriate threshold value to derive a threshold value that is not normalized.

The classification probability calculation unit 203 is configured by, for example, a probability calculation unit for a plurality of cell types. The scalars corresponding to the categories are input to the probability calculation unit for the individual categories corresponding to the categories. The probability operators for the plurality of monomer classes are each a function of the invertible transform. These functions may be different from each other or the same. The Function of the reversible transformation may be a differentiable Function, such as a sigmoid Function or a Hyperbolic Tangent Function (Tanh). The 1 st transformation may be a computation by a probability function of a reversible transformation, or may be a transformation using a database corresponding to a computation by a probability function of a reversible transformation. The database may be a table in which inputs and outputs (converted values) are associated with each other, such as a Lookup table, i.e., a Lookup table.

The classification threshold determination unit 204 determines the 2 nd classification threshold based on the classification probability values of the plurality of cell types calculated by the classification probability calculation unit 203. More specifically, the classification threshold determination unit 204 acquires classification probability values of the plurality of cell classes, and determines the 2 nd classification threshold based on the classification result using the 2 nd classification threshold for each of the acquired classification probability values of the plurality of cell classes. For example, the classification threshold determination unit 204 reads the evaluation dataset from the storage unit 201, and determines the 2 nd classification threshold based on the correct solution data of the evaluation dataset and the classification probability values of the plurality of cell classes calculated by the classification probability calculation unit 203. That is, the classification threshold determination unit 204 determines the optimal 2 nd classification threshold based on the evaluation data set. For example, the classification threshold determination unit 204 determines the classification probability values for the plurality of monomer classes based on the ratio of FP (False Positive)/FN (False Negative) for the evaluation dataset to determine the 2 nd classification threshold.

The 2 nd classification threshold may be set to a predetermined value or may be determined according to a target accuracy set by a user. When the 2 nd classification threshold is determined according to the target accuracy, the classification threshold determination unit 204 may determine the 2 nd classification threshold so that the result obtained by applying the 2 nd classification threshold to the classification probability values of the plurality of cell classes satisfies the target threshold. The target accuracy may be set for each category or may be set in common for all categories. The details of this method will be described later in the section of the operation of the threshold calculation device.

The 2 nd classification threshold may be a value different for each of the plurality of individual classes, or may be a value identical for all of the plurality of individual classes.

The threshold value transformation unit 205 performs a 2 nd transformation from a 2 nd classification threshold value to a 1 st classification threshold value for classifying data into at least 1 of a plurality of classes, which is an inverse transformation to the 1 st transformation. In other words, the threshold value transformation unit 205 transforms the 2 nd classification threshold value into a threshold value that is not normalized (i.e., the 1 st classification threshold value). The threshold conversion unit 205 may be an inverse function of a function (for example, a probability function of reversible transformation) constituting the classification probability calculation unit 203, or may be a database corresponding to calculation of an inverse function of a probability function based on reversible transformation. The database may be a table in which inputs and outputs (inverse transformed values) are associated with each other, such as a Lookup table.

The 1 st classification threshold is used in the information processing apparatus 100a to classify data into at least 1 of a plurality of classes. The information processing apparatus 100a can perform the category classification processing of data using a threshold (1 st classification threshold) that is not normalized. Therefore, in the information processing device 100a, since the classification processing can be executed based on the feature amount extracted from the data, the normalization processing is not necessary, and the calculation amount can be reduced.

The 1 st classification threshold may be set to different values for each of the plurality of classes, or may be set to a value common to all of the plurality of classes.

The 1 st output unit 206 outputs the 1 st classification threshold. More specifically, the 1 st output unit 206 outputs the 1 st classification threshold to the 1 st acquisition unit 103 of the information processing apparatus 100a via communication. The communication may be wireless communication such as Wi-Fi (registered trademark) or Bluetooth (registered trademark), or wired communication such as Ethernet (registered trademark).

The threshold calculation device 200a may also include a training unit (not shown) for training the machine learning model. The training unit may include a storage unit (not shown) for holding a training data set. The training data set includes a set of input data and correct solution data that are stored in advance. The training unit may acquire new training data from a database disposed on a server connected via a communication network such as the internet, and update the training data set. The training unit may include a holding unit (not shown) for holding the same classification model as the 1 st calculating unit 202, and the holding unit may further hold N classification probability calculating units of the same single class as the classification probability calculating unit 203. The training unit trains the same classification model as the 1 st arithmetic unit 202 using the training data set. Further, the training unit may train a network having N single-class classification probability calculation units included in the classification probability calculation unit 203. If the training of the classification model and the network is completed, the training unit may output the trained classification model to the 1 st arithmetic unit 202, and update the 1 st arithmetic unit 202 with the trained classification model. Similarly, the training unit may output a network having the trained classification probability calculation units for N individual classes to the classification probability calculation unit 203, and update the classification probability calculation unit 203 to the trained network.

[ information processing apparatus ]

Next, the information processing apparatus 100a will be explained. The information processing apparatus 100a is an apparatus that classifies data acquired by a sensor into at least 1 of a plurality of classes, and outputs the classification result. Hereinafter, an example in which data is an image will be described. Note that data and sensors for acquiring data are described in the outline of the information processing system 300a, and therefore, the description thereof is omitted.

As shown in fig. 1, the information processing apparatus 100a includes a 1 st acquisition unit 103, a 2 nd acquisition unit 101, a 2 nd calculation unit 102, a threshold processing unit 104, and a 2 nd output unit 105.

The 1 st acquiring unit 103 acquires the 1 st classification threshold outputted from the 1 st output unit 206 of the threshold calculation device 200a via communication, and outputs the acquired 1 st classification threshold to the threshold processing unit 104. Since the communication is described above, the description thereof is omitted. The 1 st classification threshold may be stored in advance in a storage unit provided in the information processing device 100 a.

The 2 nd acquisition unit 101 acquires data from the sensor via communication. Here, the sensor is an image sensor and the data is an image. The 2 nd acquisition unit 101 outputs the acquired data to the 2 nd arithmetic unit 102. The communication may be wireless communication or wired communication. The number of sensors is not limited to 1, and for example, two or more sensors may be synchronized to acquire data. The 2 nd acquiring unit 101 may acquire data via a removable memory. The removable memory is, for example, a USB (Universal Serial Bus) memory.

The 2 nd arithmetic unit 102 is a feature extractor for extracting a feature of data, and is, for example, a machine learning model. For example, the 2 nd arithmetic unit 102 is a trained classification model. The classification model is a multi-layer neural network (DNN). The data acquired by the 1 st acquisition unit 103 is input to the 2 nd arithmetic unit 102. The 2 nd arithmetic unit 102 outputs a plurality of scalars corresponding to a plurality of categories of the data. Each scalar quantity is a feature quantity corresponding to each category of data. The output of the 2 nd arithmetic unit 102 is a value that is not normalized.

The 2 nd arithmetic unit 102 is not limited to DNN. For example, the 2 nd arithmetic unit 102 may be a feature extractor other than the DNN that uses a feature point extraction method (for example, edge extraction), a principal component analysis method, a block matching method, a sampling moll method, or the like.

The threshold processing unit 104 acquires the 1 st classification threshold outputted from the 1 st acquisition unit 103, and classifies the data into at least 1 of the plurality of classes based on the output of the 2 nd arithmetic unit 102 and the 1 st classification threshold. More specifically, the threshold processing unit 104 classifies the data into at least 1 of the plurality of categories by determining whether or not the probability value for each of the plurality of categories output from the 2 nd arithmetic unit 102 is equal to or greater than the 1 st classification threshold. The 1 st classification threshold may be set to different values for each of the plurality of classes, or may be set to a value common to all of the plurality of classes. Further, more specific operation of the threshold processing unit 104 will be described later.

The 2 nd output unit 105 outputs the classification result of the data. The 2 nd output unit 105 may output the data classification result to a presentation unit (not shown), or may output the data classification result to another device other than the information processing device 100 a. For example, the 2 nd output unit 105 causes the presentation unit to present information based on the classification result based on an operation of the user input to the input unit (not shown). The input unit is, for example, a keyboard, a mouse, a touch panel, a button, a microphone, or the like. The presentation unit is, for example, a display or a speaker. The information processing apparatus 100a may or may not include an input unit and a presentation unit. The input unit and the presentation unit may be provided by other devices than the information processing device 100a, for example. The other devices than the information processing device 100a may be information terminals such as a smart phone, a tablet computer, or a computer. Note that, the information processing apparatus 100a is exemplified by a computer, but may be provided in a server connected via a communication network such as the internet.

The information processing apparatus 100a may also include a training unit (not shown) for training the machine learning model, as in the threshold value calculation apparatus 200 a. As for the details of the training unit, in the description of the training unit described in the threshold value calculation device 200a, the training of the network including the classification probability calculation unit 203 may not be included, and the 1 st calculation unit 202 may be referred to as the 2 nd calculation unit 102 instead.

The information processing system 300a may further include a training unit (not shown) that is common to the threshold value calculation device 200a and the information processing device 100a and that trains the classification model of the 1 st calculation unit 202, the network of the classification probability calculation units 203, and the classification model of the 2 nd calculation unit 102.

[ operation of threshold calculation device ]

Next, the operation of the threshold calculation device 200a will be described with reference to fig. 2. Fig. 2 is a flowchart showing an example of the operation of the threshold calculation device 200a according to embodiment 1.

The 1 st arithmetic unit 202 reads input data of the evaluation dataset from the storage unit 201, and outputs a plurality of scalars corresponding to a plurality of categories of the input data (step S1001). The scalars corresponding to the categories of the input data are feature quantities of the input data for the categories. For example, it is assumed that the input data is an image, and a car and a motorcycle are captured in the image. Assume multiple categories such as pedestrian, car, motorcycle, bicycle, and background. In this case, the 1 st arithmetic unit 202 outputs vectors (pedestrian, automobile, motorcycle, bicycle, background) having a plurality of scalars corresponding to a plurality of categories of input data, for example, (0.1, 90, 60, 0.01, 0.001).

Next, the classification probability calculation unit 203 performs 1 st conversion, which is conversion from the output of the 1 st calculation unit 202 into classification probability values of a plurality of cell categories (step S1002). As described above, the classification probability calculation unit 203 includes a classification probability calculation unit for a plurality of cell types. The classification probability calculation unit for each cell type is a probability function of reversible transformation. The classification probability calculation units for the individual categories may be different functions or the same function. The classification probability calculation units for the individual categories may be databases corresponding to calculations performed by probability functions of reversible transformations. The threshold value calculation unit for each individual category converts the scalar quantity for each of the plurality of categories into a value in the range of 0 to 1 by reversible transformation (i.e., normalization). For example, if each scalar quantity (0.1, 90, 60, 0.01, 0.001) of (pedestrian, automobile, motorcycle, bicycle, background) which is the output of the 1 st arithmetic unit 202 is input to the classification probability arithmetic unit of the individual category corresponding to each category, each scalar quantity is converted into a value in the range of 0 to 1 in the classification probability arithmetic unit of each individual category. The classification probability calculation unit 203 outputs (pedestrian, automobile, motorcycle, bicycle, background) as the classification probability values of the plurality of cell categories as (0.3, 1.0, 1.0, 0.1, 0). Here, the normalization is not performed so that a plurality of scalar values are added to become 1, but each scalar value is converted into a value in the range of 0 to 1 in accordance with the magnitude of each scalar value. In this case, the transform coefficient may be adjusted according to the classification accuracy of each class.

Next, the classification threshold determination unit 204 sets a 2 nd classification threshold based on the classification probability values of the plurality of cell classes derived in step S1002 (step S1003). More specifically, the classification threshold determination unit 204 acquires classification probability values of the plurality of cell classes, and determines the 2 nd classification threshold based on the classification result using the 2 nd classification threshold for each of the acquired classification probability values of the plurality of cell classes. For example, the classification threshold determination unit 204 reads the evaluation dataset from the storage unit 201, and determines the 2 nd classification threshold based on the correct solution data of the evaluation dataset and the classification probability values of the plurality of cell types. For example, the classification threshold determination unit 204 may determine the 2 nd classification threshold based on a ratio of fp (false positive)/fn (false negative) with respect to the evaluation dataset for each of the classification probability values of the plurality of cell classes. Further, for example, the 2 nd classification threshold may be determined so that the result obtained by applying the 2 nd classification threshold to the outputs (i.e., classification probability values) of the plurality of cell classes satisfies the target accuracy. More specifically, a target accuracy may be set, the classification accuracy in the case where the threshold (2 nd classification threshold) of the classification probability values of the plurality of individual categories is changed may be calculated, the threshold with the calculated classification accuracy closest to the target accuracy may be selected, and the threshold may be determined as the 2 nd classification threshold. For example, instead of the threshold value at which the calculated classification accuracy is closest to the target accuracy, the threshold value at which the calculated classification accuracy first exceeds the target accuracy may be determined as the 2 nd classification threshold value. The target accuracy may be set for each category, or may be set in common for all categories, for example. The 2 nd classification threshold may be different values for each of the plurality of individual classes, or may be the same value for all of the plurality of individual classes.

Next, the threshold value transformation unit 205 performs a 2 nd transformation, which is a transformation from the 2 nd classification threshold value set by the classification threshold value determination unit 204 to the 1 st classification threshold value for classifying the data into at least 1 of the plurality of classes, and is an inverse transformation to the 1 st transformation (step S1004). This makes it possible to obtain a threshold value (here, the 1 st classification threshold value) that is not normalized from the normalized threshold value (here, the 2 nd classification threshold value). The 1 st classification threshold is a threshold used to classify data into at least 1 of a plurality of classes. The 1 st classification threshold may be set to a different value for each class, or may be set to a value common to all of the classes. For example, when the threshold conversion unit 205 is an inverse function of the function (for example, a probability function of reversible conversion) constituting the classification probability calculation unit 203, the threshold conversion unit 205 calculates the 1 st classification threshold into which the 2 nd classification threshold is inversely converted, using the 2 nd classification threshold derived in step S1003 as an input. In the case where the threshold conversion unit 205 is a database corresponding to the calculation by the inverse function of the function constituting the classification probability calculation unit 203 (for example, a table in which inputs and outputs are associated as in the Lookup table), if the 2 nd classification threshold is input to the threshold conversion unit 205, the 1 st classification threshold corresponding to the input is output.

Next, the 1 st output unit 206 outputs the 1 st classification threshold derived in step S1004 to the information processing apparatus 100a (step S1005). In this case, the 1 st output unit 206 may be communicably connected to the information processing apparatus 100 a. Since the communication method is described above, the description thereof is omitted.

[ operation of information processing apparatus ]

Next, the operation of the information processing apparatus 100a will be described with reference to fig. 3. Fig. 3 is a flowchart showing an example of the operation of the information processing apparatus 100a according to embodiment 1.

The 2 nd acquisition unit 101 acquires data from a sensor (not shown) such as an image sensor (step S2001). Here, an example in which the data is an image will be described. The image may be a moving image or a still image (also referred to simply as an image). The 2 nd acquisition unit 101 may be communicably connected to the sensor, or may acquire a plurality of images from the sensor via a removable memory, for example, a usb (universal Serial bus) memory. Since the communication method is described above, the description thereof is omitted.

The 2 nd arithmetic unit 102 outputs a plurality of scalars corresponding to a plurality of categories of the data acquired in step S2001 (step S2002). The data may be, for example, a moving image captured by an onboard camera or an image. In the case of a moving image, the following operations are performed for each of a plurality of images constituting the moving image. Like the operation of the 1 st arithmetic unit 202 of the threshold calculation device 200a, the 2 nd arithmetic unit 102 outputs a plurality of scalars (feature quantities) based on the acquired data. Hereinafter, a case where the output of the 2 nd arithmetic unit 102 is (pedestrian, automobile, motorcycle, bicycle, background) — 70, 90, 0.5, 60, 0.001 will be described.

The 1 st acquiring unit 103 acquires the 1 st classification threshold outputted from the threshold calculation device 200a (step S2003). At this time, the 1 st acquiring unit 103 acquires the 1 st classification threshold from the threshold calculation device 200a via communication. Since the 1 st classification threshold and the communication method are described above, the description thereof is omitted. Here, it is assumed that the 1 st classification threshold is (pedestrian, car, motorcycle, bicycle, background) ═ 60, 60, 60, 60, 60.

The threshold processing unit 104 classifies the data acquired by the 2 nd acquiring unit 101 into at least 1 of the plurality of classes based on the output of the 2 nd calculating unit 102 and the 1 st classification threshold acquired by the 1 st acquiring unit 103 (step S2004). More specifically, in step S2004, the threshold processing unit 104 classifies the data into at least 1 category of the plurality of categories by determining whether or not each of the plurality of scalars corresponding to the plurality of categories of the data is equal to or greater than the 1 st classification threshold.

For example, when the output of the 2 nd arithmetic unit 102 is (pedestrian, automobile, motorcycle, bicycle, background) ═ 70, 90, 0.5, 60, 0.001), and the 1 st classification threshold value of each class is (pedestrian, automobile, motorcycle, bicycle, background) ═ 60, 60, 60, 60, 60, 60), the scalar value for 3 classes is equal to or greater than the 1 st classification threshold value. In this case, the threshold processing unit 104 may classify the image into 3 categories of pedestrians, automobiles, and bicycles, or into 1 category. In the latter case, the threshold processing unit 104 may select a class indicating the largest scalar value among the probability values for 3 classes of the image, and classify the image into 1 class. In this case, the probability value for the car category is the largest among the scalar values for the 3 categories described above, so the threshold processing unit 104 classifies the image as the car category. The threshold processing unit 104 may classify the image into 1 category for a category of a scalar value having a large difference from the 1 st classification threshold among the scalar values of the 3 categories. In this case, the threshold processing unit 104 may classify the image into a car category. As described above, the data classification method may be appropriately set according to the type of data, the purpose of classification, and the like.

The 2 nd output unit 105 outputs the classification result obtained in step S2004 (step S2005). The 2 nd output unit 105 may output the classification result to a presentation unit (not shown) or may output the classification result to another device other than the information processing device 100 a. For example, the 2 nd output unit 105 causes the presentation unit to present information based on the classification result based on an operation of the user input to the input unit (not shown). Since the input unit, the presentation unit, and other devices are described above, the description thereof is omitted.

The information based on the classification result may present various types of classification results based on the setting input to the input unit. For example, in the case where the data is an image captured by an in-vehicle camera, the information based on the classification result may be the type and number of detections of the object type detected by the information processing device 100a, the recognition accuracy of each object type, a change in recognition accuracy due to weather or a time zone with respect to each object type, a tendency of an image with low recognition accuracy, a recommendation for avoiding danger, or the like. The information processing apparatus 100a may transmit the analysis result to a database disposed in a server connected via the internet, for example, and acquire information based on the analysis result.

(modification example)

Next, an information processing apparatus according to a modification of embodiment 1 will be described. Hereinafter, differences from embodiment 1 will be mainly described, and descriptions of common points will be omitted or simplified.

[ overview of information processing apparatus ]

Fig. 4 is a block diagram showing an example of the configuration of the information processing apparatus 100b according to the modification of embodiment 1. The information processing device 100b includes a threshold calculation unit 20a, a storage unit 201a, and an information processing unit 10.

In embodiment 1, an example is described in which the information processing apparatus 100a acquires the 1 st classification threshold from the threshold calculation apparatus 200a, and classifies data into at least 1 of a plurality of classes using the acquired 1 st classification threshold. The main difference between the information processing apparatus 100b and the information processing apparatus 100a according to embodiment 1 is the provision of the threshold value calculation unit 20 a. In fig. 1 and 4, substantially the same components are denoted by the same reference numerals.

In embodiment 1, as shown in fig. 1, the threshold calculation device 200a outputs the 1 st classification threshold derived by the threshold conversion unit 205 to the 1 st output unit 206, and outputs the 1 st classification threshold to the information processing device 100a via the 1 st output unit 206. In the present modification, as shown in fig. 4, the threshold value calculation unit 20a stores the 1 st classification threshold value derived by the threshold value conversion unit 205 in the storage unit 201 a. Then, the information processing unit 10 reads out the 1 st classification threshold value stored in the storage unit 201 a.

As shown in fig. 1 and 4, the acquisition unit 101a of fig. 4 corresponds to the 2 nd acquisition unit 101 of fig. 1, and the output unit 105a of fig. 4 corresponds to the 2 nd output unit 105 of fig. 1. That is, the acquisition unit 101a acquires data from a sensor such as an image sensor, as in the case of the 2 nd acquisition unit 101. Similarly to the 2 nd output unit 105, the output unit 105a outputs a classification result of classifying the data acquired by the acquisition unit 101a into at least 1 of the plurality of classes.

[ operation of threshold calculation section ]

Fig. 5 is a flowchart showing an example of the operation of the threshold value calculation unit 20a according to the modification of embodiment 1. Since steps S3001 to S3004 in fig. 5 correspond to steps 1001 to S1004 in fig. 2, respectively, the description thereof will be simplified. The operation of the threshold value calculation unit 20a according to the modification differs from the operation of the threshold value calculation device 200a according to embodiment 1 in that the 1 st classification threshold value is stored in the storage unit 201 a.

The 1 st arithmetic unit 202 reads input data of the evaluation dataset from the storage unit 201, and outputs a plurality of scalars corresponding to a plurality of categories of the input data (step S3001).

Next, the classification probability calculation unit 203 performs 1 st conversion, which is conversion of the output from the 1 st calculation unit 202 into classification probability values of a plurality of cell categories (step S3002).

Next, the classification threshold determination unit 204 sets a 2 nd classification threshold based on the classification probability values of the plurality of cell types derived in step S3002 (step S3003).

Next, the threshold conversion unit 205 performs a 2 nd conversion on the 2 nd classification threshold set by the classification threshold determination unit 204, the 2 nd conversion being a conversion from the 2 nd classification threshold to a 1 st classification threshold for classifying the data into at least 1 of the plurality of classes, and being an inverse conversion to the 1 st conversion (step S3004).

Next, the threshold conversion unit 205 stores the 1 st classification threshold in the storage unit 201a (step S3005).

[ operation of information processing Unit ]

Fig. 6 is a flowchart showing an example of the operation of the information processing unit 10 according to the modification of embodiment 1. Since steps S4001 and S4002 in fig. 6 correspond to steps S2001 and S2002 in fig. 3, the description thereof will be simplified.

The acquisition unit 101a acquires data from a sensor such as an image sensor (step S4001). Next, the 2 nd arithmetic unit 102 outputs scalar quantities (probability values) corresponding to the categories of the data acquired in step S4001 (step S4002).

Next, group processing for each of a plurality of categories of data is started.

The threshold processing unit 104 reads out the 1 st classification threshold stored in the storage unit 201a (step S4003). Here, an example in which the threshold processing unit 104 reads out the 1 st classification threshold for each of the plurality of classes in accordance with the classification processing for each class is described, but the threshold processing unit 104 may temporarily read out the 1 st classification threshold for all of the plurality of classes. The 1 st classification threshold may be different values for each of the plurality of classes or may be the same value.

The threshold processing unit 104 classifies the data acquired by the 2 nd acquiring unit 101 into at least 1 of the plurality of classes based on the output of the 2 nd calculating unit 102 and the 1 st classification threshold read out from the storage unit 201 a. First, the 1 st classification threshold of 1 of the plurality of classes (for example, the class of a pedestrian) is read from the storage unit 201a (step S4003), and it is determined whether or not the scalar value of the class is equal to or greater than the 1 st classification threshold read from the storage unit 201a (step S4004). When the scalar value of the category is equal to or greater than the 1 st classification threshold (Yes in step S4004), the threshold processing unit 104 associates the scalar value of the category with a number indicating the category and stores the associated number in the storage unit 201a (step S4005).

On the other hand, when the scalar value of the category is smaller than the 1 st classification threshold (No in step S4004), the threshold processing unit 104 reads out the 1 st classification threshold of the other 1 categories (for example, the category of the automobile) among the plurality of categories from the storage unit 201a (step S4003). The threshold processing unit 104 determines whether or not the scalar value of the category is equal to or greater than the 1 st classification threshold read from the storage unit 201a (step S4004). When the scalar value of the category is equal to or greater than the 1 st classification threshold (Yes in step S4004), the threshold processing unit 104 associates the scalar value of the category with a number indicating the category and stores the associated number in the storage unit 201a (step S4005).

On the other hand, when the scalar value of the category is smaller than the 1 st classification threshold (No in step S4004), the threshold processing unit 104 reads out 1 other category (for example, category of motorcycle) among the plurality of categories from the storage unit 201a (step S4003), and executes the determination process of step S4004. By repeating the same processing in this manner, classification processing for each category is executed for a plurality of scalars corresponding to a plurality of categories of data. If the group processing according to the category is finished, the threshold processing unit 104 determines whether or not the number of categories stored in the storage unit 201a is 1 or more (step S4006). When the number of categories stored in the storage unit 201a is 1 or more (Yes in step S4006), the threshold processing unit 104 outputs the category number having the largest scalar value to the output unit 105a (step S4007). Thus, the output unit 105a outputs a classification result (for example, data is an automobile) of classifying the data into at least 1 of the plurality of classes based on the class number (not shown).

On the other hand, when the number of categories stored in the storage unit 201a is zero (No in step S4006), the threshold processing unit 104 outputs a number indicating the others to the output unit 105a (step S4008). The others, as described, represent categories that do not conform. In this case, the classification result output from the output unit 105a may be, for example, data of another type or data of a background type.

(embodiment mode 2)

Next, an information processing system according to embodiment 2 will be described. Hereinafter, differences from embodiment 1 will be mainly described, and descriptions of common points will be omitted or simplified.

[ overview of information processing System ]

Fig. 7 is a block diagram showing an example of the configuration of the information processing system 300b according to embodiment 2. The information processing system 300b includes the threshold calculation device 200b and the information processing device 100 a. The main difference between the threshold value calculation device 200b according to embodiment 2 and the threshold value calculation device 200a according to embodiment 1 is that the 1 st classification threshold value is derived from the 2 nd classification threshold value set by the user.

[ Structure of threshold calculation device ]

Next, the configuration of the threshold calculation device 200b will be described.

As shown in fig. 7, the threshold value calculation device 200b includes an input unit 207, a threshold value conversion unit 205, and a 1 st output unit 206. Here, an example in which the threshold calculation device 200b includes the input unit 207 will be described, but the present invention is not limited thereto. For example, the input unit may be provided in a device other than the threshold calculation device 200 b. Other devices are for example tablet terminals, smart phones or computers etc.

The input unit 207 inputs an operation signal from the user to the threshold value conversion unit 205. The input unit 207 is, for example, a touch panel, a keyboard, a mouse, buttons, a speaker, or the like. The operation signal is, for example, a signal indicating a 2 nd classification threshold value for each of the plurality of classes. The user inputs the 2 nd classification threshold value to the threshold value conversion unit 205 via the input unit 207. The 2 nd classification threshold may be a preset value.

The threshold value transformation unit 205 performs a 2 nd transformation, which is an inverse transformation to the 1 st transformation, on the acquired 2 nd classification threshold value, and derives a 1 st classification threshold value. The 1 st output unit 206 outputs the 1 st classification threshold to the information processing apparatus 100 a.

Note that the information processing apparatus 100a is the same as the information processing apparatus 100a according to embodiment 1, and therefore, the description thereof is omitted.

[ operation of threshold calculation device ]

Next, the operation of the threshold calculation device 200b will be described. Fig. 8 is a flowchart showing an example of the operation of the threshold calculation device 200b according to embodiment 2.

Although not shown, the user inputs the 2 nd classification threshold for each of the plurality of classes to the threshold value conversion unit 205 via the input unit 207.

As shown in fig. 8, the threshold conversion unit 205 acquires the 2 nd classification threshold of a plurality of classes input via the input unit 207 (step S5001). Next, the threshold conversion unit 205 performs a 2 nd transform on the acquired 2 nd classification threshold, the 2 nd transform being a transform from the 2 nd classification threshold to a 1 st classification threshold for classifying the data into at least 1 of the plurality of classes, and being an inverse transform to the 1 st transform (step S5002).

The 1 st output unit 206 outputs the 1 st classification threshold to the information processing apparatus 100a (step S5003).

Note that the operation of the information processing device 100a is the same as the example described in embodiment 1, and therefore the description thereof is omitted.

(modification example)

Next, an information processing apparatus according to a modification of embodiment 2 will be described. Hereinafter, differences from embodiment 2 will be mainly described, and descriptions of common points will be omitted or simplified.

[ overview of information processing apparatus ]

Fig. 9 is a block diagram showing an example of the configuration of an information processing apparatus 100c according to a modification of embodiment 2. The information processing device 100b includes a threshold calculation unit 20b, a storage unit 201b, and an information processing unit 10.

In embodiment 2, an example has been described in which the information processing apparatus 100a acquires the 1 st classification threshold from the threshold calculation apparatus 200b, and classifies data into at least 1 of a plurality of classes using the acquired 1 st classification threshold. The main difference between the information processing apparatus 100c according to the present modification and the information processing apparatus 100a according to embodiment 2 is that the information processing apparatus 100c includes a threshold value calculation unit 20 b. The main difference between the threshold value calculation unit 20b according to the present modification and the threshold value calculation device 200b according to embodiment 2 is that the threshold value calculation unit 20b includes a classification threshold value determination unit 204. In fig. 7 and 9, substantially the same components are denoted by the same reference numerals.

In embodiment 2, as shown in fig. 7, the threshold calculation device 200b inversely transforms the 2 nd classification threshold inputted to the threshold transformation unit 205 via the input unit 207 by the threshold transformation unit 205 to derive the 1 st classification threshold. Then, the threshold calculation device 200b outputs the derived 1 st classification threshold from the 1 st output unit 206 to the information processing device 100. In the present modification, as shown in fig. 9, the threshold calculation unit 20b stores the 1 st classification threshold derived by the threshold conversion unit 205 in the storage unit 201 b. Then, the information processing unit 10 reads out the 1 st classification threshold stored in the storage unit 201b, and classifies the data into at least 1 of the plurality of classes.

As shown in fig. 7 and 9, the acquisition unit 101a of fig. 9 corresponds to the 2 nd acquisition unit 101 of fig. 7, and the output unit 105a of fig. 9 corresponds to the 2 nd output unit 105 of fig. 7. That is, the acquisition unit 101a acquires data from a sensor such as an image sensor, as in the case of the 2 nd acquisition unit 101. Similarly to the 2 nd output unit 105, the output unit 105a outputs a classification result of classifying the data acquired by the acquisition unit 101a into at least 1 of the plurality of classes.

As shown in fig. 1 and 9, the classification threshold determination unit 204 in fig. 9 corresponds to the classification threshold determination unit 204 in fig. 1. The classification threshold determination unit 204 determines an appropriate 2 nd classification threshold by referring to the evaluation data set stored in the storage unit 201 or the storage unit 201b, for example.

In the threshold value calculation device 200b according to embodiment 2, the 1 st classification threshold value is derived from the 2 nd classification threshold value input by the user, but in the threshold value calculation unit 20b according to the present modification, the 2 nd classification threshold value is determined based on the 2 nd classification threshold value input by the user and the evaluation data stored in the storage unit 201 b. The details of the determination process will be described in the item of operation.

The information processing unit 10 is the same as the information processing unit 10 according to the modification of embodiment 1, and therefore, the description thereof is omitted.

[ operation of threshold calculation section ]

Fig. 10 is a flowchart showing an example of the operation of the threshold value calculation unit 20b according to the modification of embodiment 2.

Although not shown, the user inputs the 2 nd classification threshold for each of the plurality of classes to the classification threshold determination unit 204 via the input unit 207. Thus, as shown in fig. 10, the classification threshold determination unit 204 acquires the input 2 nd classification threshold (step S6001).

Next, the classification threshold determination unit 204 reads the evaluation data from the storage unit 201b (step S6002). The classification threshold determination unit 204 determines whether the class 2 threshold acquired in step S6001 is appropriate, more specifically, whether the result obtained by applying the class 2 threshold to the classification probability values of the plurality of individual categories in the evaluation dataset satisfies the target accuracy, based on the evaluation dataset (step S6003). If it is determined that the 2 nd classification threshold is appropriate (Yes in step S6003), the classification threshold determination unit 204 outputs the 2 nd classification threshold to the threshold conversion unit 205. The threshold value transformation unit 205 performs a 2 nd transformation from the 2 nd classification threshold value to a 1 st classification threshold value for classifying the data into at least 1 of the plurality of classes, and is an inverse transformation to the 1 st transformation, on the acquired 2 nd classification threshold value (step S6004). The threshold conversion unit 205 stores the derived 1 st classification threshold in the storage unit 201b (step S6005).

On the other hand, if it is determined that the 2 nd classification threshold value is not appropriate (No in step S6003), the classification threshold value determination unit 204 causes a presentation unit (not shown) to present a message that the 2 nd classification threshold value is not appropriate (step S6006). At this time, the classification threshold determination unit 204 may present, to the presentation unit, a 2 nd classification threshold that satisfies the target accuracy as a result of applying the 2 nd classification threshold to the classification probability values of the plurality of cell types of the evaluation data set.

The operation of the information processing unit 10 is the same as that of the information processing unit 10 according to the modification of embodiment 1, and therefore, the description thereof is omitted.

(other embodiments)

Although the information processing apparatus and the information processing method according to 1 or more embodiments have been described above based on the embodiments, the present invention is not limited to these embodiments. The present invention is not limited to the embodiments described above, and various modifications and combinations of the embodiments described above may be made without departing from the spirit and scope of the present invention.

For example, the processing described in the above embodiments may be realized by centralized processing using a single device (system), or may be realized by distributed processing using a plurality of devices. The processor that executes the program may be single or plural. That is, the collective processing may be performed, or the distributed processing may be performed.

In addition, the above-described embodiments may be variously modified, replaced, added, omitted, and the like within the scope of the claims and the equivalents thereof.

Industrial applicability

The present invention is useful as an information processing device, an information processing method, a program, and the like, which can reduce the amount of computation for classifying the types of objects.

Description of the reference symbols

10 information processing part

20a, 20b threshold value calculating part

100a, 100b, 100c information processing apparatus

101 No. 2 acquisition unit

101a acquisition unit

102 2 nd arithmetic operation part

103 1 st acquisition part

104 threshold processing unit

105 2 nd output part

105a output unit

200a, 200b threshold value calculating device

201. 201a, 201b storage unit

202 1 st arithmetic unit

203 classification probability calculation unit

204 classification threshold determination unit

205 threshold value conversion unit

206 output part 1

207 input unit

300a, 300b information processing system

Claims

1. An information processing device is provided with a processor,

the processor performs the following processing:

obtaining a 1 st classification threshold for classifying the data into at least 1 of a plurality of classes;

outputting a classification result of classifying the data into at least 1 of the plurality of classes based on an output of a trained classification model and the 1 st classification threshold;

the 1 st classification threshold is obtained by a 2 nd transform of a 2 nd classification threshold, the 2 nd transform being an inverse transform to the 1 st transform;

the 1 st transformation is a transformation of classification probability values from an output of the trained classification model to a plurality of individual classes constituting the plurality of classes;

the 2 nd classification threshold is set based on the classification probability values of the plurality of cell types.

2. The information processing apparatus according to claim 1,

the output of the trained classification model is a plurality of scalars corresponding to the plurality of classes.

3. An information processing method, executed by a computer,

performing a 1 st transformation of classification probability values from the output of the trained classification model to a plurality of monomer classes;

setting a 2 nd classification threshold value based on the classification probability values of the plurality of monomer classes;

performing a 2 nd transform, said 2 nd transform being a transform from said 2 nd classification threshold to a 1 st classification threshold used to classify the data into at least 1 of the plurality of classes, being an inverse transform to the 1 st transform;

and outputting the 1 st classification threshold value.

4. The information processing method according to claim 3,

the 1 st transformation is an operation based on a probability function that can be inversely transformed;

the 2 nd transformation is an operation based on an inverse function of the probability function.

5. The information processing method according to claim 3,

the 1 st transformation is a transformation using a database corresponding to an operation based on a probability function that can be inversely transformed;

the 2 nd transformation is a transformation using a database corresponding to an operation based on an inverse function of the probability function.

6. The information processing method according to claim 3,

acquiring a data set;

inputting the data set to the trained classification model, and obtaining classification probability values of the plurality of cell types for each data included in the data set;

the 2 nd classification threshold is determined based on the classification result using the 2 nd classification threshold for each of the acquired classification probability values of the plurality of cell classes.

7. In a program for executing a program,

an information processing method for causing a computer to execute:

the 1 st transformation is a transformation of classification probability values from the output of the trained classification model to a plurality of cell classes;